Software SushiMultilingual Relevance Reranker

Available on AWS Marketplace · Version 1.0.0

Release status

Version 1.0.0 is published on AWS Marketplace. The product owner approved the separately scoped original-model relevance offer and documented sources/limitations on 15 November 2026. The catalog and delivery are Public, the offer is Released, and the anonymous storefront has been verified.

Gate Current evidence
Native public human relevance diagnostics 1,517 full queries; 1,343 admitted, 174 unsupported pools retained
Original-only same-stack reload 1,343/1,343 unchanged semantic responses after API metadata changes
Final immutable image sha256:506dd8f807a30844a98a03d0bc52efddf392c2971088063efaf7d4939b2ec85c
Image security Zero HIGH/CRITICAL Trivy findings; separate NVIDIA userspace review; AWS host patching not inferred
Retained failed AWS attempts Default driver 470/CUDA incompatibility; missing serving main guard; both corrected and retested
Actual real-time native qualification A10G ml.g5.xlarge: 1,343/1,343 responses and 7/7 rejection probes; frozen portability gate passed
Sequential real-time latency Client round-trip p50 136 ms, p95 205 ms after startup; no capacity or SLA claim
Actual model package certification AWS Completed, CertifyForMarketplace=true; three fictional EN/ES/PT outputs validated
Actual native batch qualification L40S ml.g6e.xlarge: 1,343/1,343 outputs; frozen portability gate passed
Data handling Network isolation enabled; native passage texts remain private and are absent from image, Git and media
Temporary qualification compute Endpoint deleted, native batch Completed, temporary guard/model removed; refreshing Terraform plan reports no changes
Videos English 1080p walkthrough and vertical Short rendered on SER9; complete decode, exported frames and independent encoded-audio ASR checked
Final scope/sources/offer Approved 15 November 2026: original relevance, sources/limitations, USD 0.50/instance-hour and 14-day software trial; AWS separate
Public Marketplace availability Public product and delivery, Released offer; verified 15 November 2026

Earlier adapted complete-answer experiments failed their unchanged goals and remain separate. Public human relevance diagnostics do not establish customer workflow accuracy, answer completeness or missing-answer detection. Native customer-workflow acceptance and independent cross-language quality remain unmeasured.

The catalog lists 16 regions. Runtime qualification and the notebook are scoped to us-east-1; availability in the catalog does not establish tested operation, instance quota or capacity in every listed region. No buyer subscription was purchased during this seller-side verification. After subscribing, use the regional ARN supplied by Marketplace configuration, as described in the notebook.